[HIP] Support flush denorms bitcode

[ashi1]

Summary

We need to use oclc_daz_opt_on.amdgcn.bc bitcode when option fcuda-flush-denormal-to-zero is enabled for HIP.

[b-sumner]

LGTM

[scchan]

LGTM

[yaxunl]

I think we should have a lit test for linked bitcode files.

lib/Driver/ToolChains/HIP.cpp

78

you can use StringRef here. Please use a more meaningful name, e.g. FlushDenormalControlBC.

[ashi1]

I've changed the string to a more meaningful name, and also added a lit test to check that oclc_daz_opt_on.amdgcn.bc is used.

[ashi1]

This comment was removed by ashi1.

[ashi1]

Fix diff file

[yaxunl]

test/Driver/hip-toolchain-flush-denorms.hip

6 ↗

(On Diff #153127)

we only need test with one arch

7 ↗

(On Diff #153127)

better put the dummy bc files under test/Driver/Inputs/hip_dev_lib

8 ↗

(On Diff #153127)

we also need to test the case without this option

[yaxunl]

lib/Driver/ToolChains/HIP.cpp

85

we need to fix the order. should be ocml, then ockl, then irif. but should do it in another patch

[ashi1]

I've updated. Moved the bitcode files into hip_dev_lib and added test for both on and off flush denorms to zero option.

[yaxunl]

better merge these two tests and rename it as hip-device-libs.hip. We can also add other device lib tests to it later.

[ashi1]

Merged checks for flush denorm on and off into single test hip-device-libs.hip

[yaxunl]

LGTM. Thanks!

lib/Driver/ToolChains/HIP.cpp

const char *AMDGCN::Linker::constructLLVMLinkCommand(
 
   // If --hip-device-lib is not set, add the default bitcode libraries.
   if (BCLibs.empty()) {
     // Get the bc lib file name for ISA version. For example,
     // gfx803 => oclc_isa_version_803.amdgcn.bc.
     std::string ISAVerBC =
         "oclc_isa_version_" + SubArchName.drop_front(3).str() + ".amdgcn.bc";
 
+    llvm::StringRef FlushDenormalControlBC;
+    if (Args.hasArg(options::OPT_fcuda_flush_denormals_to_zero))
+      FlushDenormalControlBC = "oclc_daz_opt_on.amdgcn.bc";
+    else
+      FlushDenormalControlBC = "oclc_daz_opt_off.amdgcn.bc";
+
     BCLibs.append({"opencl.amdgcn.bc",
                    "ockl.amdgcn.bc", "irif.amdgcn.bc", "ocml.amdgcn.bc",
                    "oclc_finite_only_off.amdgcn.bc",
-                   "oclc_daz_opt_off.amdgcn.bc",
+                   FlushDenormalControlBC,
                    "oclc_correctly_rounded_sqrt_on.amdgcn.bc",
                    "oclc_unsafe_math_off.amdgcn.bc", ISAVerBC});
   }
   for (auto Lib : BCLibs)
     addBCLib(C, Args, CmdArgs, LibraryPaths, Lib);
 
   // Add an intermediate output file.
   CmdArgs.push_back("-o");

test/Driver/hip-device-libs.hip

+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: amdgpu-registered-target
+
+// Test flush-denormals-to-zero enabled uses oclc_daz_opt_on
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN:   -x hip --cuda-gpu-arch=gfx900 \
+// RUN:   --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN:   -fcuda-flush-denormals-to-zero \
+// RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
+// RUN: 2>&1 | FileCheck %s --check-prefixes=COM,FLUSHD
+
+// Test flush-denormals-to-zero disabled uses oclc_daz_opt_off
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN:   -x hip --cuda-gpu-arch=gfx900 \
+// RUN:   --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN:   %S/Inputs/hip_multiple_inputs/b.hip \
+// RUN: 2>&1 | FileCheck %s --check-prefixes=COM,NOFLUSHD
+
+
+// COM: [[LLVM_LINK:"*.llvm-link"]]
+// FLUSHD-SAME: {{.*}} "{{.*}}oclc_daz_opt_on.amdgcn.bc"
+// NOFLUSHD-SAME: {{.*}} "{{.*}}oclc_daz_opt_off.amdgcn.bc"
+// COM-SAME: {{.*}} "-o" "{{.*}}-gfx900-linked-{{.*bc}}"
+